A new AI world model learns modular arithmetic by rotating numbers in its head - and it actually works on operations it has never seen.
Researchers built BRo-JEPA, a JEPA-style world model that treats handwritten digits or letters (from MNIST and EMNIST) as states and modular arithmetic operations as actions. Instead of learning each operation as a separate lookup, its block-rotation predictor represents every operation as a rotation in latent space, mirroring the cyclic structure of modular math itself. The model was trained only on the simplest operations, adding or subtracting 1, then tested on operations it never saw during training. On that zero-shot test, BRo-JEPA with a ResNet-18 encoder hit 99.44% accuracy on MNIST and 94.35% on EMNIST letters, compared with 54.54% and 25.13% for the best standard block-rotation baseline.
That gap is the actual finding. Standard supervised and JEPA models with operation embeddings look fine on operations they've already seen but collapse when asked to extrapolate, which is a proxy for a bigger question in AI: do these systems learn the underlying rule, or just memorize enough surface pattern to fake it. Building the algebraic structure directly into how the model represents transformations, rather than hoping it emerges from data, is what closed that gap here.
It's worth being clear about scope: this is digit and letter recognition plus toy arithmetic, not general reasoning, and the trick relies on knowing in advance that the task has cyclic structure to bake into the architecture. Real-world problems rarely come with that kind of structural hint attached.